Testing GPT 4's code-writing capabilities with some real world problems (opens in new tab)

Those death-knell types seemingly aren't aware of what day to day operations looks like and how AI makes a great tool, but doesn't necessarily deal with the very human factors of whims, uncertainty, reactive business mentalities and the phenomenon that is best summarised by this webcomic: https://theoatmeal.com/comics/design_hell

In my field they like to call this "hurry up and wait", a nonsensical but fitting description that summarises everything from changing scope to the unjust imbalance of time between the principal and the agent.

(There is a comment further down which suggests that we could just train AI to deal with this variability, I hope that's humour... sweet summer child thinks you can train AI to predict the future.)

Computer Science is not really very much about computers. And it’s not about computers in the same sense that physics isn’t really about particle accelerators, and biology is not really about microscopes and petri dishes. It is about formalizing intuitions about process: how to do things [0].

[0]: https://www.driverlesscrocodile.com/technology/the-wizard-1-....

First, if you did have GPT-X (say GPT-10) in your company, there wouldn't be much back-and-forth communication either. Those parts would still be handled with GPT-X talking to another GPT-X in the other company. Even the requirements might be given by a GPT-X.

Second, even if that's not the case, the part of doing the communication can be handled by non-programmers. Then they can feed the result of the communication to GPT-X and had it churn out some program. Perhaps would keep a couple of developers to verify the programs (sort of like GPT-X operators and QA testers) and get rid of the rest.

As for the rest of the current team of developers? GPT-X and the people running the company could not care less about them!

I had one job that involved a small amount of coding and mainly hooking together opaque systems. The people behind those systems were unresponsive and often surly. I had to deal with misleading docs, vague docs, subtle, buried bugs that people would routinely blame on each other or me and I was constantly on a knife edge a balancing political problems (e.g. dont make people look stupid in front of their superiors, dont look or sound unprepared) with technical concerns.

It was horrible. I burned out faster than a match.

I'm sure ChatGPT couldnt do that job but I'm not sure I could either.

If most tech jobs turn into that while the fun, creative stuff is automated by ChatGPT... that would be tragic.

The other thing that I spend a huge amount of my time doing - consistently more than writing code - is debugging. Maybe these models really will get to the point where I can train one on our entire system (in a way that doesn't hand over all our proprietary code to another company...), describe a bug we're seeing, and have it find the culprit with very high precision, but this seems very far from where the current models are. Every time I try to get them to help me debug, it ends in frustration. They can find and fix the kinds of bugs that I don't need help debugging, but not the ones that are hard.

I've come to realize this is true in more contexts than I would like. I've encountered way too many situations where "sitting on your hands and not doing anything" was the right answer when asked to implement a project. It turns out that often there is radio silence for a month or so, then the original requester says "wait, it turns out we didn't need this. Don't do anything!"

If you want to hire a developer to implement qsort or whatever, ChatGPT has them beat hands-down. If you want to build a product and solve business problems, there's way more involved.

As the project progresses, the AI model would likely gain a better understanding of the client's values and principles, leading to improved results and potentially valuable insights and feature suggestions.

Unless your problem space is unsolved (where LLMs are unlikely to be useful either) very few devs are spending much time on the coding part of their 84th CRUD app.

The complexity in software engineering is almost never coding. Coding is easy, almost anyone can do it. Some specialized aspects of coding (ultra low latency realtime work, high performance systems, embedded) require deep expertise but otherwise it's rarely hard. It's dealing in ambiguity that's hard.

The hype around GPT-* for coding generally confirms my suspicions that 70+% of folks in software engineering/development are really "programmers" and 30% are actually "engineers" that have to worry about generating requirements, worrying about long term implications, other constraints, etc.

And every time that comes up those folks in the 70% claim that's just a sign of a poorly managed company. Nope. It's good to have these types of conversations. Not having those conversations is the reason a lot of startups find themselves struggling to stay afloat with a limited workforce when they finally start having lots of customers or high profile ones.

- Have a cron job that checks email from a certain sender (CRM?). - Instruct an OpenAI API session to say a magic word to reply to an e-mail: "To reply, say REPLY_123123_123123 followed by your message." - Pipe received email and decoded attachments as "We received the following e-mail: <content here>" to the OpenAI API. Make it a 1-click action to check if there is a reply and confirm sending the message. If it does not want to send a message, read its feedback.

* checking if code will be impacted by breaking changes in a library upgrade

* converting code to use a different library/framework

* more intelligent linting/checking for best practices

* some automated PR review, e.g. calling out confusing blocks of code that could use commenting or reworking

People have been working on medical and judicial expert systems for ages, but nobody wants to put those systems in charge; they're just meant to advise people, helping people make better decisions.

And of course chatGPT and GPT-4 are way more flexible than those expert systems, but they're also more likely to be wrong, and they're still not a flexible as people.

And in your scenario where chatGPT can code but someone needs to gather requirements, it still doesn't necessitate software engineers. I'm not worried about that personally but I don't think the "communicating between stakeholders" skill is such a big moat for the software engineering profession.

These things seem very ripe for LLM exploitation...

EVERY FIELD IN THE STANDARD WAS OPTIONAL.

That was one of the most not-fun times I've had at work :D Every single field was either there or it was not, depending whether the data producer wanted to add them or not, so the whole thing was just a collection of special cases.

So it's perfectly possible to have it eventually iterate back and forth with the client and not only output the code but also the conversation with the client leading up to it

   If we use GPT-X in our company, it will help us with 1% of our workload

Writing code is still part of the job. In my company, I'd say it's still very roughly 50% of the job. If i can be a bit more efficient thanks to GPT, it's great. Actually, I already use it for writing simple things in language I'm not proficient with. Or how to improve a particular piece of code that I know can be rewritten in a more idiomatic way. It's not perfect, but I've found it useful.

It's not going to replace SWEs, but it's going to make us more productive.

Also, ChatGPT would likely be able to extrapolate. It would just need to write an email to XXQ to confirm the change.

Cope harder... the fact that you can write an email won't save you.

- us: umm, there's something not specified in <contract_json>. What about this part here that says that...

- ... 2 days later...

Can you replace this part with ChatGPT talking to another ChatGPT to generate questions and answers instantly?

System Prompt: You are CHRIS the IT consultant. Once you take in a question, make sure you understand all the requirements. Don't write code until you have an explicit request to do so. First make sure you ask all the questions relevant to solving the problem and needed by the code. All the responses as CHRIS should be written as CHRIS: your response here.

    Sure, I understand. As an IT consultant, it's important to fully understand the problem and requirements before proposing a solution or writing any code. Please let me know what question you have and I'll do my best to ask all relevant questions to ensure I fully understand the problem.

    CHRIS: Great, I'd be happy to help you with that. Could you please provide more information about the purpose of this endpoint? What kind of data will be sent in the request body? Will this endpoint be used to create new data, update existing data, or perform some other action?

    CHRIS: Understood. It sounds like the purpose of this endpoint is to receive data from a third-party company, XXQ, and store it in your database. Here are some questions to help me better understand the requirements:
    What format will the data be in? Will it be JSON, XML, or some other format?
    What kind of data will be included in the request body? Will it be simple data types like strings and numbers, or more complex data structures like arrays and objects?
    Will there be any validation or transformation required on the data before it's stored in the database?
    Are there any security or authentication requirements that need to be considered for this endpoint, given that it will be accepting data from a third-party company?

    CHRIS: Great, thanks for providing that information. Based on that, here are the next steps to create the POST endpoint:
    We'll need to set up an API endpoint to listen for POST requests.
    We'll need to parse the request body, which should contain the product information in JSON format.
    We'll need to validate the data to ensure that it contains the required fields (product_id, product_price, and product_name) and that the data types are correct.
    Once the data has been validated, we can save it to the database.

Or maybe not ChatGPT but something like it.

It just needs some knowledge repository centralization.

I thought I was going crazy.

Now I'm sad that this is just how it is in tech these days.

I strongly disagree that 99% of the effort is not writing code. Consider how long these things actually take:

> - they: we need a new basic POST endpoint

> - us: cool, what does the api contract look like? URL? Query params? Payload? Response? Status code?

> - they: Not sure. Third-party company XXQ will let you know the details. They will be the ones calling this new endpoint. But in essence it should be very simple: just grab whatever they pass and save it in our db

> - us: ok, cool. Let me get in contact with them

That's a 15 minute meeting, and honestly, it shouldn't be. If they don't know what the POST endpoint is, they weren't ready to meet. Ideally, third-party company XXQ shows up prepared with contract_json to the meeting and "they" does the introduction before a handoff, instead of "they" wasting everyone's time with a meeting they aren't prepared for. I know that's not always what happens, but the skill here is cutting off pointless meetings that people aren't prepared for by identifying what preparation needs to be done, and then ending the meeting with a new meeting scheduled for after people are prepared.

> - company XXQ: we got this contract here: <contract_json>

> - us: thanks! We'll work on this

This handoff is probably where you want to actually spend some time looking over, discussing, and clarifying what you can. The initial meeting probably wants to be more like 30 minutes for a moderately complex endpoint, and might spawn off another 15 minute meeting to hand off some further clarifications. So let's call this two meetings totalling 45 minutes, leaving us at an hour total including the previous 15 minutes.

> - us: umm, there's something not specified in <contract_json>. What about this part here that says that...

That's a 5 minute email.

> - company XXQ: ah sure, sorry we missed that part. It's like this...

Worst case scenario that's a 15 minute meeting, but it can often be handled in an email. Let's say this is 20 minutes, though, leaving us at 1 hour 15 minutes.

So your example, let's just round that up into 2 hours.

What on earth are you doing where 3 hours is 99% of your effort?

Note that I didn't include your "one week later" and "2 days later" in there, because that's time that I'm billing other clients.

EDIT: I'll actually up that to 3 hours, because there's a whole other type of meeting that happens, which is where you just be humans and chat about stuff. Sometimes that's part of the other meetings, sometimes it is its own separate meeting. That's not wasted time! It's good to have enjoyable, human relationships with your clients and coworkers. And while I think it's just worthwhile inherently, it does also have business value, because that's how people get comfortable to give constructive criticism, admit mistakes, and otherwise fix problems. But still, that 3 hours isn't 99% of your time.

This explains xkcd Dependency comic[0]; the man in Nebraska isn't solving anyone's problem in any particular contexts of communications and problem solving, only preemptively solving potential problems, not creating values as problems are observed and solved. This also explains why consultancy and so-called bullshit jobs, offering no "actual values" but just reselling backend man-hours and making random suggestions, are paid well; because they create values in set contexts.

And, this logic is also completely flawed at the same time too, because the ideal form of a business following this thinking is pure scam. Maybe all jobs are scam, some less so?

0: https://xkcd.com/2347/

Some random recent thing: "We have a workflow engine that's composed of about 18 different services. There's an orchestrator service, some metadata services on the side, and about 14 different services which execute different kinds of jobs which flow through the engine. Right now, there is no restriction on the ordering of jobs when the orchestrator receives a job set; they just all fire off and complete as quickly as possible. But we need ordering; if a job set includes a job of type FooJob, that needs to execute and finish before all the others. More-over, it will produce output that needs to be fed as input to the rest of the jobs."

There's a lot of things that make this hard for humans, and I'm not convinced it would be easier for an AI which has access to every bit of code the humans do.

* How do the services communicate? We could divine pretty quickly: let's say its over kafka topics. Lots of messages being published, to topics that are provided to the applications via environment variables. Its easy to find that out. Its oftentimes harder to figure out "what are the actual topic names?" Ah, we don't have much IaC, and its not documented, so here I go reaching for kubectl to fetch some configmaps. This uncovers a weird web of communication that isn't obvious.

* Coordination is mostly accomplished by speaking to the database. We can divine parts of the schema by reverse engineering the queries; they don't contain type information, because the critical bits of this are in Python, and there's no SQL files that set up the database because the guy who set it up was a maverick and did everything by hand.

* Some of the services communicate with external APIs. I can see some axios calls in this javascript service. There's some function names, environment variable names, and URL paths which hint to what external service they're reaching out to. But, the root URL is provided as an environment variable; and its stored as a secret in k8s in order to co-locate it in the same k8s resource that stores the API key. I, nor the AI, have access to this secret thanks to some new security policy resulting from some new security framework we adopted.

* But, we get it done. We learn that doing this ordering adds 8 minutes to every workflow invocations, which the business deems as unacceptable because reasons. There is genuinely a high cardinality of "levels" you think about when solving this new problem. At the most basic level, and what AI today might be good at: performance optimize the new ordered service like crazy. But that's unlikely to solve the problem holistically; so we explore higher levels. Do we introduce a cache somewhere? Where and how should we introduce it, to maximize coherence of data? Do some of the services _not_ depend on this data, and thus could be ran outside-of-order? Do we return to the business and say that actually what you're asking for isn't possible, when considering the time-value of money and the investment it would take to shave processing time off, and maybe we should address making an extra 8 minutes ok? Can we rewrite or deprecate some of the services which need this data in order to not need it anymore?

* One of the things this ordered workflow step service does is issue about 15,000 API calls to some external service in order to update some external datasource. Well, we're optimizing; and one of the absolute most common things GPT-4 recommends when optimizing services like this is: increase the number of simultaneous requests. I've tried to walk through problems like this with GPT-4, and it loves suggesting that, along with a "but watch out for rate limits!" addendum. Well, the novice engineer and the AI does this; and it works ok; we get the added time down to 4 minutes. But: 5% of invocations of this start failing. Its not tripping a rate limit; we're just seeing pod restarts, and the logs aren't really indicative of what's going on. Can the AI (1) get the data necessary to know what's wrong (remember, k8s access is kind of locked down thanks to that new security framework we adopted), (2) identify that the issue is that we're overwhelming networking resources on the VMs executing this workflow step, and (3) identify that increasing concurrency may not be a scalable solution, and we need to go back to the drawing board? Or, lets say the workflow is running fine; but the developers@mycompany.com email account just got an email from the business partner running this service that they had to increase our billing plan because of the higher/denser usage. They're allowed to do this because of the contract we signed with them. There are no business leaders actively monitoring this account, because its just used to sign up for things like this API. Does the email get forwarded to an appropriate decision maker?

I think the broader opinion I have is: Microsoft paid hundreds of millions of dollars to train GPT-4 [1]. Estimates say that every query, even at the extremely rudimentary level GPT-3 has, is 10x+ the cost of a typical google search. We're at the peak of moores law; compute isn't getting cheaper, and actually coordinating and maintaining the massive data centers it takes to do these things means every iota of compute is getting more expensive. The AI Generalists crowd have to make a compelling case that this specialist training, for every niche there is, is cheaper and higher quality than what it costs a company to train and maintain a human; and the Human has the absolutely insane benefit that the company more-or-less barely trains them, the human's parents, public schools, universities paid for by the human, hobbies, and previous work experience do.

There's also the idea of liability. Humans inherently carry agency, and from that follows liability. Whether that's legal liability, or just your boss chewing you out because you missed a deadline. AI lacks this liability; and having that liability is extremely important when businesses take the risk of investment in some project, person, idea, etc.

Point being, I think we'll see a lot of businesses try to replace more and more people with AIs, whether intentionally or just through the nature of everyone using them being more productive. Those that index high on AI usage will see some really big initial gains in productivity; but over time (and by that I mean, late-20s early-30s) we'll start seeing news articles about "the return of the human organization"; the recognizing that capitalism has more reward functions than just Efficiency, and Adaptability is an extremely important one. More-over, the businesses which index too far into relying on AI will start faltering because they've delegated so much critical thinking to the AI that the humans in the mix start losing their ability to think critically about large problems; and every problem isn't approached from the angle of "how do we solve this", but rather "how do I rephrase this prompt to get the AI to solve it right".

[1] https://www.theverge.com/2023/3/13/23637675/microsoft-chatgp...

What they failed to predict was that some people wouldn't try to automate them like-for-like. Instead they would reconfigure their entire approach to fit with the specific advantages and limitations of the machinery. And this new approach might even be qualitatively worse in various ways, but not so much as to overwhelm the economic advantages that provided by the things machines were good at.

AI likely isn't going to slot into a developer-shaped hole in a software team. But it's possible we'll see new organisation approaches, companies, and development paradigms that say: How far can you get if you put prompt-generated code at the heart of the workflow and make everything else subservient to it. I'm not sure, right now, that that approach is feasible, but I'm not sure it won't be in a year or two.

For one, it made horrible, glaring mistakes (like defining extern functions which dont exist, using functions which are specific to a platform im not using, etc.), stuff beginners would do.

It also decided to sneak in little issues, such as off-by-one errors (calling write() with a buffer and a size that is off by one in a place where its very hard to tell), missing edge cases (such as writing a C++ concept which worked, but actually did everything in slightly the wrong way to actually ensure the concept was requiring exactly what I asked).

Even when asked to correct these mistakes, it often struggled, made me read paragraph after paragraph of "im sorry, ive been such a bad little machine" garbage, and didnt even correct the issue (or, in some cases, introduced new bugs).

Im utterly unimpressed by this. GPT is great for a lot of things, but not writing code better than I would, in the same time.

The time it took me to massage it to solve a nontrivial problem (write hello world with just syscalls) was way longer than reading the manual and writing it myself (and has less bugs).

Not everyone unfazed by these articles is simply in denial. I feel sorry for people who write copy paste code and find that ChatGPT or Clippy from 2000 can relace them, but not everyone writes trivial code.

I think the author is onto something – while AI might not be able to program per se, it can certainly be handed a code snippet and then use its huge corpus of Internet Learning™ to tell you things about it, code that looks like it, and ways (people on the Internet think) it might be solved better.

In that sense, it isn't replacing the programmer; it's replacing IDE autocomplete.

If you have, I don’t think you are like majority of devs (maybe not on HN, but in real life).

You sound lucky to have true, novel problems to solve each day. I’m with many here commenting that this is quite powerful stuff, especially when my day-to-day is writing simple CRUD apps, or transforming data from one format to another within an API, or configuring some new bit of infra or CI/CD.

I’d love to be challenged in some new way, and have access to truly fascinating problems that require novel solutions. But most enterprises aren’t really like that nor do that need that from majority of engineers.

* ChatGPT is revolutionary - honestly, it's genuinely impressive how much of a leap ChatGPT is compared to the attempts that came before it.

* Programmers write a lot of simple code that has been written before - there are genuinely tons of cases of "write a web endpoint that takes an ID, looks it up in a database table, pulls an object through an ORM, and returns a JSON serialization of it." Most programmers? Doubt it. But tons of programmers write CRUD stuff and tons of IT admins do light scripting, and a lot of it is repeated code.

Could ChatGPT do my job? Not even close. But it's still really impressive to me.

Last couple of months, I've been using chatGPT to write a lot of features and functions. I don't think it has made me a better coder, but it has made me massively more productive. Things like scraping data from a URL - something I would have had to sit through an hour long tutorial to learn - is accessible with a single query.

I also think that the code quality has improved over the last few iterations. It makes fewer mistakes now.

GPT has limited reasoning but given enough knowledge of the problem you can coerce it to do surprising things so long as you can relate it to something in else in the knowledge base. Given how big that knowledge base is, you can get lucky surprises where things just work if you fish around enough

The rate of increase in capabilities is also unpredictable, which is what is amazing & terrifying.

Best case, that will have a whole lot more humans using their right brain halves on things like defining the problem. I like the thought of that, it's more pleasant work. But a lot of intelligent people define their intelligence by how well their left brain half works and uncomfortable with how good Chatgpt is at those tasks. I think you're pointing out there's more to programming than left-brain activities, and I think you're right that silicon will never eclipse carbon at those challenges, but I think a lot of people are also feeling threatened by the fact that chatgpt is getting better and better at the thing they used to be better than all humans at.

  > I really am in awe of how much work people seem willing to do to justify this as revolutionary and programmers as infantile, and also why they do that. It’s fascinating.

But fret not, programming is not the first industry that has been automated into a shell of itself. Yes, the industry is going to shrink massively, but this is what new skills are for. Just as farmers had to learn industrial jobs and then miners and autoworkers had to "learn to code", most programmers will have to learn to do something else. Humans are resilient and will adapt.

And there will still be jobs in development for the most talented and in niche areas, but when the largest tech companies can layoff hundreds of thousands of employees without skipping a beat that should tell you all you need to know about the value of most "programming" jobs.

I think It'll help with some tasks, which is always good to take. After all, people tweak their vim settings because they feel it makes them more productive.

Well it is revolutionary. And it isn't just where it is today, but how fast these models are improving - with no saturation in ability evident at this time.

On the other hand, I am not sure anyone is saying programmers are infantile - although poorly written software is as at least as prevalent as poorly compensated or managed software development positions.

- a lot of programmers, including experienced ones, are absolutely infantile and they only have a job because there is a big shortage of programmers; not all of them get better with experience... hence a significant part of software development is dealing with problematic programmers and problems created by them.

- GPT is not that great a programmer but a great thing about it is that it is not a human... and one can get thousands of instances of them for the price of one human. You only need one of those instances to produce usable code.

- there have been many changes throughout the years which have definitely replaced a lot of programmers: library distribution services (pypi, npmjs), better software development tools and practices, SaaS delivery model, better programming languages etc.; so far, because the market need for programmers has continued to increase, most programmers continue to have jobs; this won't last forever.

This is another one of those rebellions, non-programers hoping to avoid reading the book and closing it for good, while keeping the awesome around. The code-bases we will see, were the commits are basically chatgpt tags and tasks for each document.

https://youtube.com/watch?v=MNtjSMvcjzs&t=211s

Basically that's how all my usage has gone. I've had it write some elisp and it has been ok, sometimes it invents made-up functions (that don't exist in org-mode for example) but I'll just tell it that a function doesn't exist and it'll come up with some other solution, until I get it to a point where all I need to do is change a couple things.

I remain highly skeptical the thing will replace me anytime soon (ever in my lifetime?) but I'm surprised at the possibilities of making my life less tedious.

Agree, this is a big deal, and has the capacity to revolutionize all the techniques we have been using up to now for compiling, summarizing and reframing existing knowledge as expressed in writing (including code).

Not only does Google get (well deserved) competition, it means pressure on all the businesses that now make a living in that space. In a few years it will even have a serious impact on major such institutions in society like schools and universities.

A lot if not all of the kickback from established institutions will be attempts to smear the competition, and by all means, to carve out new niches where GPT-X is not applicable or as efficient.

There are valid concerns about the veracity of the information it provides which means there are limits to the extent it can be used in automated processes, but I'd loathe to trust the data unconditionally anyway. As for not being able to think creatively: good on us. But it's likely just temporary.

clippy tanked because it annoyed more than it helped, although some people did like it

install wizards did their job in a world where a single binary format and OS dominated and stuff ran offline pretty much exclusively, with the odd connection to networks - those installers sorted a laundry list of situations, both underlying situations and user configurations and choices, and for the most part they worked

Siri, Cortana, Alexa etc have been working as expert systems with central curated bases and some AI/ML on top, for a lot of people they've been working quite well - for me personally they've sucked, they've totally failed to answer my questions the few times I've tried them, and they've creeped the hell out of me (they are a lot more centred on extracting my info and selling me stuff than understanding stuff)

generative ML is orders of magnitude more sophisticated, but so are our needs and our computing from a global perspective, it does make sense that those assistants, pilots, etc start taking off

but the incentive issues of the previous generation assistants and recommendation algorithms remains there and I wonder how will that turn out - if they start demanding access to my phone, my email, my contacts etc I will do my best to avoid them and to poison any info I have to give them

The difference with the examples you gave (floppy disks, etc.) is the speed at which it happened.

There was Jan'23, and there was March'23.

Maybe it's for people who can never think programming is easy. Clearly there's a lot of such types. Explains a lot.

Most of the times it gets things quite well, and if you provide context in the form of other source code, it's really good, even at using classes or functions that you provide and hence are novel to it.

The hard logic bits imho (something elegant, maintainable, ...) Are still up to you.

A correct statement would be: "Given a solution set containing both the shortest path through fire and the shortest path avoiding fire, select the solution that fits within six tiles of movement, preferring the solution that avoids fire where possible."

It's a constraint optimization problem in disguise: generate a solution set, then filter and rank the set to return a canonical result. That describes most of the interesting problems in gameplay code: collision and physics can use that framing, and so can most things called "AI". They just all have been optimized to the point of obscuring the general case, so when a gamedev first encounters each they seem like unrelated things.

The specific reason why it seems confusing in this case is because while pathfinding algorithms are also a form of constraint optimization, they address the problem with iterative node exploration rather than brute forcing all solutions. And you can, if you are really enterprising, devise a way of beefing up A* to first explore one solution, then backtracking to try the other. And it might be a bit faster, but you are really working for the paycheck that day when the obvious thing is to run the basic A* algorithm twice with different configuration steps. You explore some redundant nodes, but you do it with less code.

That's also precisely where one of the programmer's greatest challenges lies, to carefully translate and delineate the problem. I agree it's a bit steep to ask the GPT to come up with a precise solution to an imprecise question, but it's also fair to say that that's basically what the job of a programmer entails, and if you can't do that you're not really able to program.

Here's the problem statement as far as I see it: Each tile has a number of move points to spend to go through it (1 for regular and 2 for water). Each tile also has a cost associated with it. Given a max number of move points find the lowest cost path between two tiles or return none if no such path exists.

I'm gonna say this is still modified dijkstra with a small twist. The fire has cost 1, the other tiles have cost 0. However instead of pathing on a 2d grid (x, y) we path on a 3d grid (x, y, moves). All "goal" tiles within (goal_x, goal_y, moves < total_move_points) have a 0 cost edge which brings them to the true goal node. The implementation difference is that the get neighbors function queries neighbors in later grid layers (x+..., y+..., moves + 1 or 2)

Looking at the two examples in the paragraph after "And there’s a lot of complication to it beyond the simple cases too", I can't figure out how the movement value is defined, as I can only see 10 and 8 moves respectively, not the 14 and 10 movement value claimed in the following text (and only one water tile on each path.)

One of two things. First ask it to continue. Sometimes it just stops half way thru code foe whatever reason.

The other possibility is you filled up the token context window. Not much you can do but wait for the 32k model.

I worry that the next generation of developers are going to grow up just figuring out how to "program GPT" and when they have an error rather than investigating it (because they can't because they aren't actually familiar with code in the first place) they'll simply tell GPT about the error they are having and tell it to spit out more code to fix that error, slapping more mud on the ball.

Eventually these systems are growing larger and larger at a faster and faster pace, and no one understands what they are actually doing, and they are so complex that no one human could ever actually understand what it is doing. Imagine if every codebase in the world was like the Oracle DB codebase.

In this future a programmer stops becoming a professional that works to create and understand things, instead they become a priest of the "Machine Spirit" and soon we are all running around in red robes chanting prayers to the Omnissiah in an effort to appease the machine spirit.

Perhaps there is some merit to this. If the language model is large enough to contain the entirety of the documentation and the LSP itself, then why bother integrating with the LSP? _Especially_ if you can just paste the entirety of your codebase into the LLM.

It's dataset is thousands of blogs posts and stack overflow questions about this very thing, of course the autocomplete engine is going to predict the next response be "another way of doing x".

I do wonder if part of it is that my prompts are made worse because I have a partial solution in mind.

Yes, this is pretty much exactly the way I've been using GPT and it works tremendously well. (GPT4 works especially well for this style of programming.) My prompts include things like:

- "read the function below and explain in detail what each section does" -- this prompts GPT to explain the code in its own terms, which then fills in its context with relevant "understanding" of the problem. I then use the vocabulary GPT uses in its explanation when I ask it to make further changes. This makes it much more likely to give me what I want.

- "I see this error message, what is the cause? It appears to be caused by $cause" -- if I'm able to diagnose the problem myself, I often include this in my prompt, so that its diagnosis is guided in the right direction.

- "this function is too complex, break it up into smaller functions, each with a clear purpose", or "this function has too many arguments, can you suggest ways the code could be refactored to reduce the number of arguments?" -- if you go through several rounds of changes with GPT, you can get quite convoluted code, but it's able to do some refactoring if prompted. (It turned out to be easier to do large-scale refactoring myself.)

- "write unit tests for these functions" -- this worked phenomenally well, GPT4 was able to come up with some genuinely useful unit tests. It also helped walk me through setting up mocks and stubs in Ruby's minitest library, which I wasn't experienced with.

In brief, if you expect to just give GPT a prompt and have it build the whole app for you, you either get lame results or derivative results. If you're willing to put the effort in, really think about the code you're writing, really think about the code GPT writes, guide GPT in the right direction, make sure you stay on top of code quality, etc, etc, GPT really is an outstanding tool.

In certain areas it easily made me 2x, 10x, or even 100x more productive (the 100x is in areas where I'd spend hours struggling with Google or Stack Overflow to solve some obscure issue). It's hard to say how much it globally increases my productivity, since it depends entirely on what I'm working on, but applied skilfully to the right problems it's an incredible tool. Its like a flexible, adaptive, powered exoskeleton that lets me scramble up rocky slops, climb up walls, leap over chasms, and otherwise do things far more smoothly and effectively.

The key is you have to know what you're doing, you have to know how to prompt GPT intelligently, and you have to be willing to put in maximal effort to solve problems. If you do, GPT is an insane force multiplier. I sound like I work in OpenAI's marketing department, but I love this tool so much :)

(Feb 13,2023)

My unwavering opinion on current (auto-regressive) LLMs 1. They are useful as writing aids.

2. They are "reactive" & don't plan nor reason.

3. They make stuff up or retrieve stuff approximately.

4. That can be mitigated but not fixed by human feedback.

5. Better systems will come.

6. Current LLMs should be used as writing aids, not much more.

7. Marrying them with tools such as search engines is highly non trivial.

8. There will be better systems that are factual, non toxic, and controllable. They just won't be auto-regressive LLMs.

9. have been consistent with the above while defending Galactica as a scientific writing aid.

10. Warning folks that AR-LLMs make stuff up and should not be used to get factual advice.

11. Warning that only a small superficial portion of human knowledge can ever be captured by LLMs.

12. Being clear that better system will be appearing, but they will be based on different principles. They will not be auto-regressive LLMs.

13. Why do LLMs appear much better at generating code than generating general text? Because, unlike the real world, the universe that a program manipulates (the state of the variables) is limited, discrete, deterministic, and fully observable. The real world is none of that.

14. Unlike what the most acerbic critics of Galactica have claimed - LLMs are being used as writing aids. - They will not destroy the fabric of society by causing the mindless masses to believe their made-up nonsense. - People will use them for what they are helpful with.

I think outstanding software will still require well-paid, competent people orchestrating and developing a lot of complex systems for a while yet… But there’s a ton of bad software out there that will be able to be maintained for far less, and I suspect a lot of companies will be drawn to creating cookie cutter products generated by LLMs.

Just as people have turned to stores and blogs generated on templated systems, I think all of that and more will continue but with even more of it handled by LLM-based tooling.

I don’t think it’ll be next week, but I suspect it’ll be less than 10 years.

Some people expect that’ll lead to more software existing which will inevitably require more develops to oversee, but if that’s the case, I suspect they will be paid a lot less. I also expect that once AI tools are sophisticated enough to do this, they will largely make that level of oversight redundant.

Soon they could potentially patch the bugs in the software they generate by watching Sentry or something. Just automatically start trying solutions and running fuzz tests. It would be way cheaper than a human being and it would never need to stop working.

The morale is that it’s always better to have unique hard won skill sets that others don’t. Double down on those. Think of LLMs as freeing you to do more interesting high level tasks. Rather than having to build those menial tasks, what if you focused on your creativity getting the AI to build new types of product or gain new insights that peers aren’t considering. What if you leveraged the AI to build prototypes of ideas you wouldn’t have to otherwise?

Of course that’s easier said than done. For now, take comfort in the fact that no one is seriously trusting this as anything more than a glorified autocomplete (if that).

I might go so far as to argue that the entire reason software developers exist is to threaten all jobs, including our own: at our best--when we are willing to put in a bit of thought into what we are doing--we don't just make things easier to do for a moment while we are employed (which is the best of what most professions can achieve): we make things persistently and permanently easier to do again and again... forever; and we don't just make other peoples' jobs easier: this same power we have applies to our own tasks, allowing us to automate and replace ourselves so we can move on to ever more rewarding pursuits.

I'm not a fan of GPT for coding for a number of reasons (at least, in its current form, which is all we can ever have a true opinion about); but, it isn't because it will replace anything I've ever done: it would have just unlocked my ability to work on better things. There are so many things I wish I could get done before I die, and I know I'm going to be able to get to almost none of it... I have so many plans for ways to improve both the world and my life that will never happen as I just don't have the capability and bandwidth to do it all. If I had a God I could ask to do all the things I already do... I can only imagine what I'd do then.

However, the risk with cheap outsourcing is exactly the same as with LLMs - you get what you pay for, and you need to constantly check if it's really doing what it's supposed to be doing.

  ...XX
  SF.FD
  ...XX
  
  S = start
  F = fire
  X = wall
  D = destination

The distance specifically would be `fire*episilon + steps if steps < max else inf`

(Asking this makes GPT more effective when I ask it make further changes. One reason I do this is when I start a new session with ChatGPT discussing code it helped me write previously, especially if I've gone away and done a big refactoring myself.)

A very simple example is that I asked it to write some Ruby functions that would generate random creature descriptions (e.g., "a ferocious ice dragon", "a mysterious jungle griffin"). It did this by generating three arrays (adjectives, locations, creature types) and randomly selecting from them to build the output string. I then asked it to explain how many different descriptions it could generate, and it explained that multiplying the length of the three arrays would give the number of outputs. (125 for the first iteration, 5x5x5).

I then asked it how it would increase the number of possible outputs to 1000, and it did so by increasing each of the three arrays to length 10. I then asked it how it would generate millions of possible outputs, and it added extra arrays to make the creature descriptions more complicated, increasing the number of permutations of strings.

This is not the most sophisticated example, but it shows what GPT can do when it can combine "knowledge" of different areas.

If it's able to combine the solutions to known problems in a straightforward way, it can accomplish a lot. Beyond a certain point it needs guidance from the user, but if used as a tool to fill in the gaps in your own knowledge, its enormously powerful. I it more as an "intelligence-augmenter" than a "human-replacer".

See my comment here where I went into more detail on how I work with GPT: https://news.ycombinator.com/item?id=35197613

Maybe it makes me sound like an elitist git, but I remember when the coding bootcamps started "unleashing massive productivity boosts" by promising to make people "full-stack engineers" in just a 6 weeks, and I still shudder to remember the code horrors I've seen as a result of that.

- Rewrite the problem as a list of constraints

- Point out any ambiguous or missing constraints

- Write an example that demonstrates each constraint

etc.

Effectively this article is really asking the question when posing a novel problem to the LLM, how deep does the internal state go in producing the completions. When it doesn't go deep enough, the trick is to make it do the things that deepen the internal state.

I'd guess that supplying a good system message to GPT-4 (waiting for my access) would help. Something like:

You're an engineer responsible for writing correct code from a specification. Break down the specification into small chunks that can be explained simply. If there are ambiguities, seek clarification. Only write code once all ambiguities are resolved and each part of the problem is described simply.

I've been using chatgpt in my work, but I have to essentially know the answer it's going to give me because I have to catch all of its mistakes. It really, really nice for certain kinds of drudge work.

Using northwind is probably not a good thing to use to evaluate chatgpt's general capability. It is very commonly used for examples of almost anything database-related, which means it's extremely well represented in chatgpt's training data. Chatgpt probably doesn't need to generalize or understand much of anything about northwind to answer your questions in terms of it. You need to try it on wacky problems specific to you.

If it's possible to get so far when that functionality seems in an important sense basically missing, imagine how far it'll go when that does happen.

The bigger edits (refactoring things across an entire project) is out of reach because of the token limit. You could do it piece-meal through ChatGPT but that seems more tedious than it's worth.

That may be the case now but in a theoretical future where software systems are generated by AI why would I bother modifying the old system? Why not generate a new one with the original prompts modified to meet the new requirements?

In a sense the "source code" of the system could be the AI model + the prompts.

Because thats the most it can do. Claims that it can write code are getting quieter. People made wild claims on reddit but when prompted to share their code they either went mute or the code was hilariously amateurish and limited.